How sebuzdugan/frai works
Project type and purpose
• FRAI is an open-source monorepo that ships three coordinated Node/TypeScript packages:
• frai-cli: end-user CLI for responsible AI workflows (documentation, scanning, evaluation, RAG indexing, fine-tune governance)
• frai-core: modular SDK powering all functionality (Questionnaire, Documents, Scanners, Rag, Eval, Finetune, Config, Providers)
• frai-agent: a LangChain-based conversational agent wrapping frai-core for natural-language orchestration
File structure
• Top-level: monorepo config (pnpm-workspace.yaml, turbo.json), lint/build configs, docs/, examples/
• packages/
• frai-core/src: core modules (config, documents, eval, finetune, providers, rag, scanners, questionnaire)
• frai-cli/src: CLI entry point, wiring to frai-core APIs
• frai-agent/src: agent definition, LangChain tool adapters to frai-core
• CI workflows under .github/workflows
README for additional insights
• Describes conceptual split (CLI vs. Core vs. Agent) and key workflows
• Lists core modules and high-level command flows (scan, generate docs, rag index, eval, finetune)
• Notes external dependency on OpenAI (via Providers)
Main components & interactions
a. User Interfaces
• Terminal / CI → frai-cli
• Chat UI or direct code → frai-agent
• Programmatic import → frai-core
b. Core Layers
• CLI & Agent wrappers
– CLI invokes frai-core APIs (Scanners.scanCodebase, Questionnaire.run, Documents.generate, Rag.index, Eval.run, Finetune.validate)
– Agent exposes two tools to LangChain: scan_repository & generate_responsible_ai_docs
• frai-core SDK
– Config/Providers: read .env or global config → wire up OpenAI (or other LLMs)
– Scanners: static analysis → detect AI indicators → return scan results
– Questionnaire: prompt user for feature details via Inquirer or other prompt function
– Documents: assemble checklist, model card, risk file from questionnaire answers
– Rag: split policy docs, embed via LLM → produce vector index (JSON store)
– Eval: baseline evaluation metrics on model outputs vs. references → JSON/Markdown report
– Finetune: generate templates, validate governance plans
c. External Services and I/O
• OpenAI API (or pluggable LLM provider) for AI-driven tips, embeddings, evaluations
• File System: read codebase, write generated docs, vector store, reports
• PDF Converter (optional markdown-pdf for docs export)
Architectural patterns & design principles
• Monorepo + Modular Packages: independent versioning for CLI, Core, Agent
• Layered Architecture: UI layer (CLI/Agent) → orchestration layer (frai-core API) → service layer (OpenAI) → persistence layer (file system, vector store)
• Plugin-style modules: each core capability is a self-contained module
• Separation of concerns:
• UI wrappers do no heavy lifting—they just call core
• Core modules focus solely on their domain